Training
Trains a GraspPolicy on GraspDataset batches, with DDP, checkpoints and periodic simulation evaluation.
Train normally, or save resumable single-device milestones and publish them.
train for TrainerConfig.cuda_graph: the same updates, with the step captured once as a CUDA graph.
Training-loop parameters for Trainer.
AdamW's learning rate drops 10x at two thirds of total_iters; that schedule has no field.
Seeds model initialization, dataset workers and sampling.
Simulation for periodic evaluation, built with build_env; None disables evaluation.
Skip training, load ExperimentConfig.ckpt_path and evaluate one trajectory per scene.
Evaluate on the train split instead of test, for small scene sets without test scenes.
Log to Weights & Biases, which also syncs the TensorBoard scalars.
W&B project used when track is set.
Global batch size in scenes; each DDP rank gets batch_size // world_size.
Evaluator parallelism; capped at batch_size during training.
Stop evaluation after this many data-loader batches.
Penetration depth (m) below which an eval grasp counts toward pass-rate (report-only).
Final training iteration; the loop runs iterations 0 through this value.
Worker processes for the training and evaluation data loaders.
AdamW learning rate.
AdamW weight decay.
Iterations between loss logs, and TensorBoard scalars when tracking.
Iterations between simulation evaluations; iteration 0 is skipped.
Iterations between iter_<N>.pt checkpoints, iteration 0 included.
Capture training on prepared HDF5 views as a CUDA graph; single GPU without evaluation.
Use TF32 and fused AdamW for ordinary training steps.
Clip the gradient norm before an optimizer step; zero disables clipping.
Per-scene point-cloud observations and successful grasp trajectories from a processed HDF5.
Load prepared observations and actions once for CUDA batch gathering.
Dataset loading parameters for GraspDataset.
One index is one scene, not one grasp; raw views resample points and actions on every access.
Successful action trajectories sampled per scene, with replacement when a scene has fewer.
Point-cloud points per scene, with replacement when the cloud is smaller.
Return the (num_actions, 3, D) supervision tensor; training needs it and evaluation turns it off.
Return per-point graspness and the best conditioning-point indices, for local-feature SpConv policies; needs data processed with enable_contact_extraction.
Sample only target-object points instead of the full cropped scene cloud.
Cache observations and actions in RAM; sampling stays unchanged.
Include RGB and camera poses when available.
Object-only multi-view data: fuse view 0 (the head) with up to this many other views, a random count per sample, into one cloud. 0: one random view.
With fuse_views: std of each view's rigid misalignment, meters (translation) and x10 radians (rotation).
With fuse_views: std of per-point Gaussian noise (m).
With fuse_views: probability of the perception augmentation (datasets/perception_aug.py) per sample; needs renders that keep the neighbours (extract_segments=[]) and camera poses in the processed h5.
Perception aug: fused-cloud keep fraction, log-uniform in [aug_keep_min, 1].
Perception aug: voxel dedup size (m), applied with p 0.5 (the old gate pipeline's 3 mm thinning).
Perception aug: probability of the head view alone (no wrist view).
Perception aug: probability of the RGB / MoGe depth-error model (per sample).